Creating Panoramic Images Using ORB Feature Detection and RANSAC-based Image Alignment

Abstract

This paper details the development and execution of a project to combine a series of horizontally overlapping photographs into a single panoramic image. The process involves the use of the Oriented FAST and Rotated BRIEF (ORB) feature detector and descriptor, and the Random Sample Consensus (RANSAC) algorithm for automatic image alignment. The final panoramic image is created by blending the images together seamlessly. Two methods for stitching the panorama are discussed: using translations with pre-spherically-warped input images, and using homographies to align the input images directly. Aiming at the problems of large scale rotation registration error, low registration rate and strong randomness and instability of random sampling consistency algorithm in the Rotational invariance binary description algorithm, a fast feature matching algorithm combining ORB and RANSAC is proposed. First, the feature point extraction method is optimized to eliminate the influence of feature edges. Then a simplified pyramid scale space model is constructed to improve the scale space structure of layered images, and reduce the number and layers of generated images; Then the gradient direction is used to improve the main direction extraction mode in the traditional ORB algorithm to improve the accuracy of the main direction of feature corners. Finally, the RANSAC algorithm is improved by constructing a block random sampling detection method to improve the stability of the RANSAC algorithm and the accuracy of Image registration.